The Black Hat talk was the most must watch item of AI content this year, but imo this is #
2#.
Owain manages to take far out Janus-esque intuitions about LLM persona selection and turn them into rigorous empirical science with important implications for near term AI safety.